Grid Demand Forecasting at a National Grid Operator — 4% MAPE on Day-Ahead Demand, 18% Lower Balancing Cost
Demand Forecaster + Real-Time Visualizer + Anomaly Detector delivering high-accuracy grid-demand forecasting in the renewables-dominated energy mix.
The challenge
The operator — a European national grid operator (the transmission-system-operator responsible for the country's electricity grid) — was operating a grid-demand-forecasting function whose accuracy had been progressively eroded by the structural changes in the country's energy mix. The country's renewables penetration had grown rapidly, with wind and solar now accounting for a substantial fraction of the total generation mix. The structural intermittency of renewables had increased the variability of the generation side of the grid-balance equation; meanwhile, the demand side was becoming more variable as well, driven by the growth in heat-pump electrification, electric-vehicle charging and behind-the-meter solar self-consumption.
The operator's existing demand-forecasting model — a classical statistical model with weather-overlay features developed in the early 2010s — was producing day-ahead demand-forecast accuracy in the 7-9% MAPE range, against an aspirational target in the 3-4% range. The forecast-accuracy gap was driving structural inefficiency in the operator's grid-balancing function — over-conservative balancing reserves (carrying excess balancing-energy at cost) when the forecast under-estimated demand-variability, and emergency-balancing actions (at premium cost) when the forecast under-estimated specific demand-events.
The constraints were significant. The forecasting model needed to integrate with the operator's existing grid-operations framework without disruption; the operator's grid-operations team had limited tolerance for forecasting-system changes that complicated the grid-operations workflow. The data inputs spanned multiple sources (the operator's own metering infrastructure, weather services, the distribution-network operators' demand data, the electric-vehicle-charging platforms) with structural integration complexity.
The approach
MindMap deployed a grid-demand-forecasting platform composed of Demand Forecaster (Df) repurposed for grid-demand-forecasting, Real-Time Visualizer (Rv) for the grid-operations team's forecasting-dashboard layer, Anomaly Detector (Ad) for the demand-event-detection layer, and Data Lake Architect (Dl) for the underlying multi-source data integration.
Phase one was the multi-source data integration. The platform's data-integration layer ingests from the operator's metering infrastructure (high-frequency demand data at the substation level), the weather services (high-resolution weather forecasting), the distribution-network operators' demand data (with appropriate data-sharing arrangements), the electric-vehicle-charging platforms (for the EV-charging demand component), and the public macro-and-economic data feeds.
Phase two was the forecasting-model build. The model is a multi-horizon ensemble: a temporal-fusion-transformer for the structural-demand component (capturing the seasonal, weekly and daily patterns with the macro-economic and weather-conditions modulation), gradient-boosted-tree models for the demand-component-specific patterns (heat-pump-demand modelling, EV-charging-demand modelling, industrial-demand modelling), and a meta-model that ensembles the component models into the final demand-forecast with calibrated confidence intervals.
Phase three was the grid-operations-workflow integration. The platform's forecasts feed into the operator's existing grid-operations framework — the day-ahead forecasting cycle, the intra-day forecasting refresh, the real-time forecasting nowcast. The grid-operations team's dashboard layer presents the current forecast, the confidence intervals, the recent forecast-vs-actual performance, and the recommended-balancing-actions where the forecasting indicates specific operational decisions.
Phase four was the demand-event-detection layer. Anomaly Detector identifies emerging demand patterns that the standard forecasting model has not captured — sudden behavioural-pattern shifts (e.g. holiday-period patterns differing from historical baselines), localised demand-events (industrial-demand-shifts at the distribution-network level), and emerging-pattern-detection (e.g. heat-wave-driven demand-pattern shifts at the early stages before the model has had time to incorporate them).
The pre-built building blocks
Rather than commission a ground-up build, the engagement leaned on MindMap's pre-built accelerator library — production-tested components that compress what would otherwise be a six-to-nine-month build into weeks.
Demand Forecaster
Multi-horizon grid-demand forecasting ensemble
Real-Time Visualizer
Grid-operations team dashboard layer
Anomaly Detector
Demand-event-detection and emerging-pattern surfacing
Data Lake Architect
Multi-source data integration with per-source data-quality monitoring
The architecture
The platform runs on the operator's private cloud with full local data-residency. The grid-demand data — operationally sensitive and with regulatory data-handling requirements — stays inside the operator's perimeter.
The data-integration layer uses Kafka for the streaming data and Snowflake for the analytics warehouse. The data volume is approximately 1.4 million data points per minute at peak. The data-integration layer handles the data-quality variations across the multiple sources (the metering infrastructure data is high-quality; the third-party data feeds have variable quality) with per-source data-quality monitoring.
Demand Forecaster's multi-horizon ensemble runs scheduled forecasting cycles aligned with the operator's grid-operations cycles — day-ahead forecasting at the standard market-cycle timing, intra-day forecasting on the operator's intra-day refresh cadence, real-time nowcasting on a 5-minute cadence. The model serving uses the operator's CPU infrastructure for the production forecasting (the model size and inference latency support CPU-based serving; the cost-efficiency outweighs the GPU-based serving alternative).
Real-Time Visualizer provides the grid-operations team's dashboard layer, designed around the team's existing operational workflow. The dashboards surface the actionable forecasting-information without overwhelming the user, with the role-based view supporting the different grid-operations roles (the system controllers, the balancing-team, the trading-desk, the operational-planning team).
Anomaly Detector runs continuously on the streaming demand-data, with the emerging-pattern-detection feeding into the forecasting model's re-calibration and into the grid-operations team's situational-awareness.
Integration with the operator's existing grid-operations framework uses the operator's standard inbound APIs, with the platform's forecasts flowing into the existing framework rather than replacing it.
The numbers behind the story
Day-ahead demand-forecasting MAPE has dropped to approximately 4% from the pre-platform 7-9% baseline. The intra-day forecasting accuracy has improved correspondingly, with the platform's 5-minute nowcast supporting the operator's real-time grid-operations decisions with materially better demand-signal than the previous model provided.
System-balancing cost has dropped approximately 18% on the operator's measurement. The reduction is driven by the combination of less-conservative balancing-reserves (the better forecast accuracy supports tighter reserve-margin operations) and reduced emergency-balancing actions (the better demand-event detection reduces the surprises that drive premium-cost balancing actions). The cost-reduction translates into a meaningful saving for the country's electricity-consumers, who ultimately bear the balancing-cost through grid-tariffs.
The grid-operations team's situational-awareness has improved. The Anomaly Detector layer surfaces emerging demand-patterns that the previous forecasting model had not been able to flag, allowing the grid-operations team to engage with emerging conditions earlier in the operational cycle.
Renewables-integration outcomes have improved. The better demand-forecasting combined with the operator's existing renewables-generation-forecasting has improved the overall grid-balance forecasting, with the consequent improvements in renewables-curtailment optimisation (the operator can integrate more renewables-generation without proportional curtailment) and in the operator's ability to support the country's continued renewables expansion.
An unexpected outcome: the platform's demand-component models (the heat-pump-demand modelling, the EV-charging-demand modelling) have produced insights that the operator's strategic-planning team has used in the network-investment-planning work. The demand-pattern projections for the country's electrification scenarios have informed the operator's transmission-network-investment decisions for the coming decade.
“Our previous demand-forecasting model had been progressively eroded by the structural changes in our energy mix. MindMap's platform delivered four-per-cent day-ahead MAPE and eighteen-per-cent lower system-balancing cost, with the augmentation-pattern integration our grid-operations team required. The cost-reduction translates directly into lower grid-tariffs for our country's electricity consumers.”— Chief Operating Officer· National Grid Operator
Why MindMap was chosen
The operator had evaluated two specialist energy-forecasting vendors. Both had strong domain expertise but limited willingness to integrate with the operator's existing grid-operations framework — both proposed wholesale-replacement approaches that the grid-operations team considered unacceptable.
MindMap's accelerator-composition approach — bringing Demand Forecaster, Real-Time Visualizer, Anomaly Detector and Data Lake Architect together with the augmentation-pattern integration (preserving the existing grid-operations framework and adding the forecasting capability on top) — was the structural differentiator.
Our embedded energy-sector expertise on the delivery team (two former TSO grid-operations leads from peer European grid operators and a former demand-forecasting modeller) was the third factor. The operator's COO felt that the team understood the operational reality of grid operations in the renewables-dominated context, not just the forecasting technology.
Related deployments
Utility Billing Disputes
ChatNext + DocuMage automated 78% of billing-dispute resolution to same-day — at a UK energy utility under intense regulatory pressure on customer service.
Wealth Advisor Co-Pilot
An LLM-grounded co-pilot for wealth advisors — meeting prep, post-meeting synthesis, compliance drafting — freeing 31% more client-facing time.
Commercial Underwriting Automation
DocuMage + Multi-Agent Orchestrator automated the data-capture layer of commercial underwriting — 4x quote throughput at the same underwriter headcount.
Want an outcome like this?
Start with a 2-week AI Readiness Sprint. We deliver a prioritised use-case backlog and business case grounded in what's actually buildable with our accelerator library.